An entity grammar-based multi-paradigm AI rule fusion method and system
By transforming multi-paradigm models into a unified structure using entity grammar, the problem of inconsistent rules in multi-paradigm model fusion is solved, and stable and traceable reasoning results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Filing Date
- 2026-05-22
- Publication Date
- 2026-07-10
AI Technical Summary
In existing technologies, multi-paradigm AI models suffer from inconsistent rules and difficulties in identifying conflicts during fusion, leading to unstable inference results and a lack of traceability. This is especially true in scenarios where knowledge graphs, causal inference, probabilistic inference, and neural symbol rules are all involved, making effective fusion and verification difficult.
Using entity grammar, rule records from different paradigm models are converted into a unified five-field structure. Through entity alignment, rule normalization, and arbitration, cross-paradigm composite rules are identified and generated, enabling unified comparison and fusion of rules.
It achieves rule consistency fusion of multi-paradigm models, reduces error propagation, improves the stability and credibility of inference results, and provides a traceable fusion inference process.
Smart Images
Figure CN122366437A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-paradigm AI rule fusion method and system based on entity grammar. Background Technology
[0002] With the development of artificial intelligence (AI) technology, various paradigms such as knowledge graph reasoning, neural symbolic reasoning, causal inference, probabilistic inference, multimodal feature fusion, and meta-learning have been widely applied to complex task processing. Each of these paradigms has its own distinct advantages; for example, knowledge graphs are suitable for expressing entity relationships and domain knowledge, causal inference is suitable for describing the causal influence between variables, probabilistic inference is suitable for handling uncertainty problems, and multimodal fusion is suitable for integrating heterogeneous information such as text, images, and numerical data. However, in practical applications, the same target task often requires the simultaneous invocation of multiple AI paradigm models for joint processing. Existing solutions typically employ interface splicing or result-level fusion, simply summarizing the outputs of different models or re-inputting them into downstream models for processing. While this approach can achieve multi-model collaboration to some extent, the data units, rule forms, constraint relationships, and reasoning paths used within different paradigm models are not uniform. This makes it difficult to directly compare rules between models, detect conflicts in a timely manner, and the fused reasoning results lack clear rule sources and processing basis. Especially in scenarios involving knowledge graph rules, causal rules, probability rules, and neural symbol rules, different rules may have problems such as similar antecedents but opposite consequents, overlapping applicable conditions but inconsistent reasoning conclusions, and inconsistent statistical associations and causal directions. If we rely solely on human experience or simple weight fusion, it will not only easily lead to unstable reasoning results, but also make it difficult to trace and verify the fusion process.
[0003] To address this, a multi-paradigm AI rule fusion method and system based on entity grammar is proposed. Summary of the Invention
[0004] The main objective of this invention is to provide a multi-paradigm AI rule fusion method and system based on entity grammar, which can effectively solve the problems in the background technology.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A multi-paradigm AI rule fusion method based on entity grammar includes the following steps:
[0007] S1. Obtain at least two preset artificial intelligence paradigm models for the same target task and target task data. The preset artificial intelligence paradigm model is a model that can output rule records or can be converted to generate rule records. The rule records include at least rule antecedents and rule consequents.
[0008] S2. Extract basic data units, abstract combined objects, operation functions, constraint rules and initial task configurations from each preset artificial intelligence paradigm model, and write the extraction results into the corresponding entity syntax five-field structure;
[0009] S3. Normalize the entity fields and rule fields in the five-field structure of different entity syntaxes to form a cross-normalization entity index table and a normalization rule set;
[0010] S4. Based on the normalized rule set, perform rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models to obtain cross-paradigm rule conflict results.
[0011] S5. Based on the cross-paradigm rule conflict results, the conflict rule records are arbitrated to obtain the retention rules, the downgraded rules, and the candidate rules.
[0012] S6. In the normalized rule set after arbitration, identify the rule combinations that satisfy the cross-paradigm association conditions, and generate cross-paradigm composite rules based on the rule combinations.
[0013] S7. Write the retention rule, weight reduction rule and cross-paradigm composite rule into the fusion rule set, and perform reasoning on the target task data based on the fusion rule set, and output the fusion reasoning result and the rule source record, conflict handling record and derivation path record corresponding to the fusion reasoning result.
[0014] Furthermore, the preset artificial intelligence paradigm model includes at least two of the following: knowledge graph reasoning model, neural symbol reasoning model, causal inference model, probabilistic inference model, multimodal feature fusion model, and meta-learning model;
[0015] Specifically, for knowledge graph reasoning models, entities, relations, triples, graph traversal operations, and ontology reasoning rules are extracted; for neural symbolic reasoning models, neural feature vectors, symbolic propositions, mapping functions between neural features and symbolic propositions, and symbolic reasoning rules are extracted; for causal inference models, causal variables, causal edges, intervention operations, observation operations, and causal constraint rules are extracted; for probabilistic inference models, random variables, conditional probability distributions, sampling operations, probability update operations, and probabilistic reasoning rules are extracted; for multimodal feature fusion models, text features, image features, numerical features, modal alignment functions, and cross-modal consistency rules are extracted; and for meta-learning models, task sets, meta-parameters, task adaptation functions, and cross-task transfer rules are extracted.
[0016] Furthermore, the entity syntax five-field structure includes a basic data unit field, an abstract entity field, an operation function field, a rule field, and an initial configuration field;
[0017] Specifically, input variables that cannot be further decomposed, feature vectors, atomic propositions, triples, or random variables are written into the basic data unit field; semantic entities, causal chains, compound propositions, probabilistic models, cross-modal semantic objects, or task adaptation objects formed by combining basic data units are written into the abstract entity field; functions used to transform, combine, align, infer, sample, update, or adapt basic data units or abstract entities are written into the operation function field; rule records including rule antecedents, rule consequents, rule sources, rule confidence, and rule application conditions are written into the rule field; and target task data, initial observation data, initial symbolic proposition set, prior distribution, domain knowledge base, or multimodal input data are written into the initial configuration field.
[0018] Furthermore, in step S3, when normalizing the entity fields in the five-field structure of different entity syntaxes, an entity alignment score is calculated based on the entity name, entity type, hierarchical relationship, and entity vector representation. The entity alignment score satisfies the following:
[0019]
[0020] in, Indicates the first The entity and the first Entity alignment score between entities Indicates the first The entity and the first Name similarity between entities Indicates the first The entity and the first Type consistency between entities Indicates the first The entity and the first The degree of matching of hierarchical relationships between entities Indicates the first The entity and the first Vector proximity between entities , , , These are the corresponding weight coefficients, and ;
[0021] When the entity alignment score is greater than the preset alignment threshold, the first... The entity and the first Multiple entities are written to the same cross-paradigm entity index entry.
[0022] Furthermore, in step S4, when performing rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models, the rule conflict strength between two rule records is calculated, and the rule conflict strength satisfies:
[0023]
[0024] in, Indicates the first Rule record and number The intensity of rule conflicts between rule records Indicates the first Rule record and number Similarity of rule antecedents between rule records Indicates the first Rule record and number The mutual exclusion of rule consequents between rule records. Indicates the first Rule record and number The degree of overlap in the applicable conditions of the rules between the rule records. , , These are the corresponding weight coefficients, and ;
[0025] When the rule conflict intensity is greater than the preset conflict threshold, determine the first... Rule record and number These rule records constitute cross-paradigm rule conflicts.
[0026] Furthermore, the cross-paradigm rule conflict includes at least one of the following: conflict of the same antecedent but different consequent, conflict of the antecedent being contained but the consequent being opposite, conflict of mutually exclusive rule consequents, conflict of overlapping rule application conditions but inconsistent reasoning results, and conflict of inconsistent causal direction and statistical correlation direction.
[0027] Furthermore, in step S5, when arbitrating conflict rule records, a rule arbitration score for the conflict rule record is calculated, and the rule arbitration score satisfies:
[0028]
[0029] in, Indicates the first The rule arbitration score of the conflict rule record. Indicates the first The rule confidence of each conflict rule record. Indicates the first The accuracy of the conflict rules recorded in historical target tasks. Indicates the first Each conflict rule record shows the degree of matching with the current target task scenario. Indicates the first Each conflict rule record corresponds to a rule source weight. Indicates the first The intensity of conflict between a conflict rule record and other conflict rule records. , , , , These are the corresponding weight coefficients;
[0030] Conflict rule records with the highest arbitration score and whose score difference with other conflicting rule records is greater than the first scoring threshold are identified as retained rules; conflict rule records with arbitration scores lower than retained rules but higher than the candidate threshold are identified as downgraded rules; conflict rule records with a difference between rule arbitration scores less than the second scoring threshold are added to the candidate rule pool; conflict rule records whose processing results cannot be determined by automatic arbitration are generated as pending review records.
[0031] Furthermore, the conflict resolution record includes conflict rule identifier, rule source tag, rule antecedent, rule consequent, rule applicable conditions, rule conflict intensity, rule arbitration score, resolution result, and review status;
[0032] When the verification accuracy of a candidate rule in the candidate rule pool reaches the threshold for candidate rule conversion in subsequent target task verification, the corresponding candidate rule is written into the fusion rule set; when the weighted rule falls below the weighted retention threshold in a preset number of consecutive target task verifications, the corresponding weighted rule is moved from the fusion rule set into the candidate rule pool.
[0033] Furthermore, in step S6, generating cross-paradigm composition rules includes:
[0034] Identify the entity association, causal association, semantic similarity, or task dependency relationship between the rule consequent of the first rule record and the rule antecedent of the second rule record;
[0035] When the association score corresponding to the entity association, causal association, semantic similarity, or task dependency is greater than the composite rule generation threshold, the first rule record and the second rule record are determined as candidate combined rules.
[0036] Candidate composite rules are generated based on the candidate combination rules. The composite antecedent of the candidate composite rule includes at least the rule antecedent of the first rule record and the rule antecedent of the second rule record. The composite consequent of the candidate composite rule includes at least the rule consequent of the second rule record.
[0037] The candidate composite rules are subjected to consistency verification of predecessors and successors, verification of rule applicability conditions, and verification of derivation validity.
[0038] When the consistency verification of predecessor and successor, the rule applicability verification, and the derivation validity verification all pass, the candidate composite rule is determined as a cross-paradigm composite rule and written into the fusion rule set;
[0039] A fusion derivation path set is generated based on the fusion rule set. The fusion derivation path set is compared with the single paradigm derivation path set corresponding to each preset artificial intelligence paradigm model to obtain new derivation paths that exist only in the fusion derivation path set. The new derivation path is then associated with the rule record it calls, the rule source tag, the conflict handling record, and the fusion reasoning result.
[0040] A multi-paradigm AI rule fusion system based on entity grammar includes:
[0041] The model access module is used to acquire at least two preset artificial intelligence paradigm models for the same target task and target task data. The preset artificial intelligence paradigm models are models that can output rule records or can be converted to generate rule records. The rule records include at least rule antecedents and rule consequents.
[0042] The component mapping module is used to extract basic data units, abstract combined objects, operation functions, constraint rules and initial task configurations from each preset artificial intelligence paradigm model, and write the extraction results into the corresponding entity syntax five-field structure;
[0043] The normalization module is used to normalize the entity fields and rule fields in the five-field structure of different entity syntaxes, forming a cross-normalization entity index table and a normalization rule set;
[0044] The rule conflict handling module is used to perform rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models according to the normalized rule set, to obtain cross-paradigm rule conflict results, and to arbitrate the conflicting rule records according to the cross-paradigm rule conflict results to obtain retained rules, reduced weight rules and candidate rules.
[0045] The composite rule generation module is used to identify rule combinations that satisfy cross-paradigm association conditions in the normalized rule set after arbitration processing, and generate cross-paradigm composite rules based on the rule combinations.
[0046] The fusion inference output module is used to write the retention rules, weight reduction rules and cross-paradigm composite rules into the fusion rule set, and to perform inference on the target task data based on the fusion rule set, and output the fusion inference result as well as the rule source record, conflict handling record and derivation path record corresponding to the fusion inference result.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] 1. This solution converts multiple pre-defined AI paradigm models targeting the same objective task into a five-field entity syntax structure. This allows the basic data units, abstract combined objects, operation functions, constraint rules, and initial task configurations in different paradigm models to be organized in a unified field format, avoiding the inconsistency in rule expression caused by simple merging at the output layer in existing multi-model splicing methods. At the same time, by normalizing the entity fields and rule fields in different entity syntax five-field structures, a cross-paradigm entity index table and a normalized rule set are established. This enables knowledge graph rules, neural symbol rules, causal rules, probabilistic rules, and multimodal consistency rules to be compared and processed in the same rule space, thereby providing a unified data foundation for subsequent rule conflict identification and fusion reasoning.
[0049] 2. This invention also identifies cross-paradigm rule conflicts through rule antecedent matching and rule consequent compatibility judgment. It then arbitrates conflicting rules by combining rule confidence, historical verification accuracy, target task scenario matching, rule source weight, and conflict intensity. This allows conflicting rules to be distinguished as retained rules, downgraded rules, and candidate rules, reducing error propagation caused by rule inconsistencies in joint inference between different paradigm models. Furthermore, this invention can identify rule combinations that satisfy cross-paradigm association conditions within the normalized rule set after arbitration and generate cross-paradigm composite rules. This ensures that different paradigm models are not merely adding up results, but forming verifiable combined inference chains at the rule level. By outputting fusion inference results, rule source records, conflict resolution records, and derivation path records, this invention enables the fusion inference process to have good traceability and auditability, facilitating improved stability and credibility of inference results in application scenarios requiring multi-model collaboration, such as medical auxiliary analysis, scientific knowledge discovery, complex decision support, and multimodal question answering. Attached Figure Description
[0050] Figure 1 This is a flowchart of the method of the present invention;
[0051] Figure 2 This is a system module diagram of the present invention;
[0052] Figure 3 This is a schematic diagram illustrating the rule fusion and conflict arbitration of this invention. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0054] Example 1
[0055] like Figure 1-3 As shown, this embodiment provides a multi-paradigm AI rule fusion method based on entity grammar, which can be deployed on servers, industrial computers, cloud inference platforms, or local intelligent analysis devices. This method is suitable for scenarios requiring simultaneous invocation of multiple artificial intelligence paradigm models for joint inference, such as medical auxiliary analysis, scientific knowledge discovery, complex fault diagnosis, multimodal question answering, and intelligent decision support. For ease of explanation, the following description uses the example of a "knowledge graph inference model, causal inference model, and probabilistic inference model" jointly participating in the same target task; however, this invention is not limited to this combination and can be replaced or extended to a neural symbolic inference model, a multimodal feature fusion model, a meta-learning model, etc.
[0056] In practical implementation, the system first acquires multiple preset artificial intelligence paradigm models and target task data for the same objective task. The target task data can be structured data, unstructured text, knowledge base data, image features, historical sample records, real-time acquired data, or a combination of the above. For example, in a medical auxiliary analysis scenario, the target task data may include patient symptom text, examination indicators, past medical history, medical knowledge graphs, and historical case data. Knowledge graph inference models are used to provide entity relationships between diseases, symptoms, examination indicators, and treatment plans; causal inference models are used to describe the causal effects between intervention factors, pathological variables, and outcome variables; and probabilistic inference models are used to infer probabilistic relationships between uncertain symptoms, examination results, and disease risks. These models can be directly accessed by the system or their rule bases, parameter files, graph data, inference logs, or intermediate output results can be read through interfaces.
[0057] In step S1, the system performs model access and rule record acquisition for each preset artificial intelligence paradigm model. For models that can directly output rule records, the system reads their rule antecedents, rule consequents, rule sources, and applicable conditions. For models that cannot directly output rule records, the system generates rule records based on the model's intermediate results, inference chains, or feature mapping relationships. The rule antecedent represents the conditions that must be met to trigger the rule, and the rule consequent represents the conclusion or state change obtained after the rule is triggered. For example, a rule record in a knowledge graph inference model can be represented as "If symptom A exists and symptom A has a graph association with disease B, then disease B is a candidate disease"; a rule record in a causal inference model can be represented as "If intervention D is performed, then variable X changes"; and a rule record in a probabilistic inference model can be represented as "When indicator C is abnormal, the posterior probability of disease B increases."
[0058] In step S2, the system extracts basic data units, abstract combination objects, operation functions, constraint rules, and initial task configurations from each preset artificial intelligence paradigm model and writes them into a five-field entity syntax structure. Specifically, for the knowledge graph reasoning model, the system uses entities, relations, and triples as basic data units, ontology concepts, entity clusters, and reasoning links as abstract combination objects, graph traversal, relation expansion, and ontology reasoning as operation functions, triple constraint rules and ontology reasoning rules as constraint rules, and the initial set of entities involved in the current target task as the initial task configuration. For the causal inference model, the system uses causal variables and observed variables as basic data units, causal chains, causal graph structures, and variable combinations as abstract combination objects, observation operations, intervention operations, and counterfactual computation as operation functions, causal direction constraints, intervention constraints, and variable dependency constraints as constraint rules, and the initial observation samples or intervention conditions as the initial task configuration. For the probabilistic inference model, the system uses random variables and conditional variables as basic data units, probability distributions, conditional probability tables and Bayesian network structures as abstract combination objects, sampling, inference and probability update as operation functions, conditional probability rules and posterior update rules as constraint rules, and prior distribution and current observation evidence as initial task configuration.
[0059] In this embodiment, the entity syntax five-field structure includes a basic data unit field, an abstract entity field, an operation function field, a rule field, and an initial configuration field. The basic data unit field stores input variables, feature vectors, atomic propositions, triples, or random variables that cannot be further decomposed or do not require further decomposition in the current task. The abstract entity field stores semantic entities, causal chains, compound propositions, probabilistic models, cross-modal semantic objects, or task-adaptive objects formed by combining basic data units. The operation function field stores functions that transform, combine, align, infer, sample, update, or adapt basic data units or abstract entities. The rule field stores rule records consisting of rule antecedents, rule consequents, rule sources, rule confidence, and rule application conditions. The initial configuration field stores target task data, initial observation data, initial symbolic proposition sets, prior distributions, domain knowledge bases, or multimodal input data. Through this field-based processing, data and rules from different paradigm models can be written into a unified structure, facilitating subsequent comparison, conflict detection, and fusion reasoning.
[0060] In step S3, the system normalizes the entity fields and rule fields in the five-field structure of different entity syntaxes. During entity normalization, the system first standardizes entity names based on a preset terminology dictionary, entity type encoding table, and domain thesaurus. For example, "hypertension," "high blood pressure," and "hypertension" can be mapped to the same standard entity item; "abnormal blood sugar" and "high fasting blood sugar" can be written into the same entity index item or parent-child entity index item based on hierarchical relationships. Subsequently, the system calculates the entity alignment score based on the entity name, entity type, hierarchical relationship, and entity vector representation. The entity alignment score can satisfy:
[0061]
[0062] in, Indicates the first The entity and the first Entity alignment score between entities Indicates the first The entity and the first Name similarity between entities Indicates the first The entity and the first Type consistency between entities Indicates the first The entity and the first The degree of matching of hierarchical relationships between entities Indicates the first The entity and the first Vector proximity between entities , , , These are the corresponding weight coefficients, and When the entity alignment score is greater than the preset alignment threshold, the system writes the corresponding entity into the same cross-paradigm entity index item; when the entity alignment score is within the manual confirmation range, the system writes the entity pair into the pending confirmation index item; when the entity alignment score is lower than the preset alignment threshold, the system retains its independent entity identifier.
[0063] During rule normalization, the system uniformly encodes the antecedent, consequent, source, confidence level, and applicability conditions of each rule record. The antecedent can be encoded as a combination of one or more condition objects, and the consequent can be encoded as a conclusion object, state change object, or probability update object. The source identifies whether the rule originates from a knowledge graph reasoning model, causal inference model, probabilistic inference model, neural symbolic inference model, or other predefined AI paradigm models. The confidence level can be output from the original model or calculated based on historical sample validation results, expert annotation levels, or rule hit stability. The applicability conditions limit the data scope, task scenario, time range, or variable conditions to which the rule applies.
[0064] In step S4, the system performs rule antecedent matching and rule consequent compatibility judgment based on the normalized rule set to identify cross-paradigm rule conflicts. For any two rule records from different preset artificial intelligence paradigm models, the system calculates the antecedent similarity, consequent mutual exclusion, and overlap of applicable conditions, and further calculates the rule conflict strength. The rule conflict strength can satisfy:
[0065]
[0066] in, Indicates the first Rule record and number The intensity of rule conflicts between rule records Indicates the first Rule record and number Similarity of rule antecedents between rule records Indicates the first Rule record and number The mutual exclusion of rule consequents between rule records. Indicates the first Rule record and number The degree of overlap in the applicable conditions of the rules between the rule records. , , These are the corresponding weight coefficients, and When the antecedents have high similarity, the consequents have high mutual exclusivity, and the applicable conditions overlap, the system determines that the two corresponding rules constitute a cross-paradigm rule conflict. The cross-paradigm rule conflict can include conflicts where the antecedent is the same but the consequent is different, conflicts where the antecedent is included but the consequent is opposite, conflicts where the rule and the consequent are mutually exclusive, conflicts where the applicable conditions of the rules overlap but the reasoning results are inconsistent, and conflicts where the causal direction is inconsistent with the statistical association direction.
[0067] For example, in a specific task, a knowledge graph inference model gives the rule "Abnormal symptoms A and indicator C both point to disease B," while a probabilistic inference model gives the rule "When indicator C is abnormal, the probability of disease B is lower." Both rules apply to the current patient data. The system then calculates the antecedent similarity and consequent mutual exclusion of the two rules. When the antecedents of both rules involve symptom A and indicator C, and the consequents point to "disease B is a candidate conclusion" and "the probability of disease B is lower," respectively, if the conflict intensity exceeds a preset conflict threshold, the two rules are recorded as a cross-paradigm rule conflict. Similarly, a causal inference model might assume that variable X has a direct causal influence on variable Y, while a statistical association rule only indicates an inverse correlation between variable Y and variable X. If their applicable conditions overlap and their inference directions are inconsistent, this can also be recorded as a conflict between causal and statistical association directions.
[0068] In step S5, the system arbitrates the conflicting rule records. Arbitration does not simply delete low-confidence rules; instead, it calculates a rule arbitration score based on rule confidence, historical verification accuracy, current target task scenario matching degree, rule source weight, and conflict intensity. The rule arbitration score must satisfy the following:
[0069]
[0070] in, Indicates the first The rule arbitration score of the conflict rule record. Indicates the first The rule confidence of each conflict rule record. Indicates the first The accuracy of the conflict rules recorded in historical target tasks. Indicates the first Each conflict rule record shows the degree of matching with the current target task scenario. Indicates the first Each conflict rule record corresponds to a rule source weight. Indicates the first The intensity of conflict between a conflict rule record and other conflict rule records. , , , , These are the corresponding weight coefficients;
[0071] The system determines the processing result of conflicting rules based on rule arbitration scores. Specifically, when a conflicting rule has the highest rule arbitration score and the score difference between it and other conflicting rules is greater than the first score threshold, the system determines that rule as a retained rule. When a conflicting rule's score is lower than that of a retained rule but higher than the candidate threshold, the system determines it as a de-weighted rule. De-weighted rules can still participate in reasoning, but their reasoning weight is lower than that of retained rules. When the score difference between multiple conflicting rules is less than the second score threshold, and the system cannot clearly determine the priority rule, the corresponding rule is written into the candidate rule pool. When a rule lacks necessary fields, its confidence level cannot be determined, or its conflict type cannot be automatically determined, the system generates a pending review record and outputs the record to the manual review interface. Each conflict processing generates a conflict processing record, which includes the conflicting rule identifier, rule source tag, rule antecedent, rule consequent, rule applicable conditions, rule conflict intensity, rule arbitration score, processing result, and review status.
[0072] In one optional implementation, the system also dynamically updates candidate rules and demotion rules. When the verification accuracy of a candidate rule in the candidate rule pool reaches the candidate-to-positive rule threshold in subsequent target task verifications, the system writes the corresponding candidate rule into the fusion rule set. When a demotion rule falls below the demotion retention threshold in a preset number of consecutive target task verifications, the system moves the corresponding demotion rule from the fusion rule set into the candidate rule pool. In this way, the system can continuously revise the fusion rule set based on the results of subsequent task executions, reducing misjudgments caused by one-time rule arbitration.
[0073] In step S6, the system identifies rule combinations that satisfy cross-paradigm association conditions in the normalized rule set after arbitration processing, and generates cross-paradigm composite rules. Specifically, the system identifies whether there is an entity association, causal association, semantic similarity, or task dependency relationship between the consequent of the first rule record and the antecedent of the second rule record. When the corresponding association score is greater than the composite rule generation threshold, the system determines the first and second rule records as candidate combination rules. Subsequently, the system generates candidate composite rules based on the candidate combination rules. The composite antecedent of the candidate composite rule includes at least the antecedent of the first rule record and the antecedent of the second rule record, and the composite consequent of the candidate composite rule includes at least the consequent of the second rule record. After generating the candidate composite rule, the system performs consistency verification of its antecedent and consequent, rule applicability verification, and derivation validity verification. Only when all the above verifications pass is the candidate composite rule determined as a cross-paradigm composite rule and written into the fusion rule set.
[0074] For example, a knowledge graph reasoning model contains the rule "symptom A is related to disease B," a causal inference model contains the rule "intervention D reduces variable X," and a probabilistic inference model contains the rule "the risk of disease B decreases when variable X decreases." When the system identifies entity or causal relationships between disease B, variable X, and intervention D, it can generate a composite rule: "When symptom A points to disease B, intervention D acts on variable X, and variable X is related to the risk of disease B, the reasoning result for the risk of disease B is corrected based on the change in variable X." This composite rule is not derived from the original rules of a single paradigm model, but is generated through cross-paradigm association based on a unified entity index and a normalized rule set, thus supporting a more complete fusion reasoning process.
[0075] In step S7, the system writes the retention rules, weight reduction rules, and cross-paradigm composite rules into the fusion rule set, and performs inference on the target task data based on the fusion rule set. During the inference process, the system calls the relevant rules in the fusion rule set according to the entity index items hit by the target task data and the rule application conditions. For retention rules, the system participates in the result calculation according to the normal inference weight; for weight reduction rules, the system participates in auxiliary inference according to the weighted rules; for cross-paradigm composite rules, the system triggers when all its composite antecedents are satisfied, and writes the composite rule consequent into the intermediate inference result or the final fusion inference result. The fusion inference result output by the system may include the target object classification result, candidate conclusion ranking, risk score, recommendation result, explanatory text, or structured decision result.
[0076] Simultaneously, the system generates rule source records, conflict resolution records, and derivation path records corresponding to the fusion inference results. The rule source record identifies which preset AI paradigm model the rules invoked by the fusion inference results originate from; the conflict resolution record explains whether the rules involved in the inference underwent conflict detection, arbitration, weight reduction, or manual review; and the derivation path record displays the rule invocation order and intermediate inference nodes from the target task data to the fusion inference results. Furthermore, the system can generate a fusion derivation path set based on the fusion rule set, and generate single-paradigm derivation path sets based on the normalized rule sets corresponding to each preset AI paradigm model. The system compares the union of the fusion derivation path set and the single-paradigm derivation path sets to obtain new derivation paths that exist only in the fusion derivation path set. A correspondence is then established between these new derivation paths and their invoked rule records, rule source tags, conflict resolution records, and fusion inference results. Thus, the system can provide traceable evidence of the fusion inference process while outputting the inference results.
[0077] This embodiment also provides a multi-paradigm AI rule fusion system based on entity grammar. The system includes a model access module, a component mapping module, a normalization processing module, a rule conflict handling module, a composite rule generation module, and a fusion inference output module. The model access module acquires at least two preset artificial intelligence paradigm models for the same target task, as well as the target task data. The component mapping module extracts basic data units, abstract composite objects, operation functions, constraint rules, and initial task configurations from each preset artificial intelligence paradigm model, and writes the extraction results into the corresponding entity grammar five-field structure. The normalization processing module normalizes the entity fields and rule fields in different entity grammar five-field structures to form a cross-paradigm entity index table and a normalized rule set. The rule conflict handling module performs rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models based on the normalized rule set, obtains cross-paradigm rule conflict results, and arbitrates conflicting rule records based on the cross-paradigm rule conflict results. The composite rule generation module identifies rule combinations that satisfy cross-paradigm association conditions in the normalized rule set after arbitration processing, and generates cross-paradigm composite rules. The fusion inference output module writes the retained rules, deweighted rules, and cross-paradigm composite rules into the fusion rule set, and performs inference on the target task data based on the fusion rule set, outputting the fusion inference results as well as the corresponding rule source records, conflict handling records, and derivation path records.
[0078] In practical deployments, the above modules can be deployed on the same server or distributed across different computing nodes. The cross-paradigm entity index table, normalized rule set, candidate rule pool, fusion rule set, and fusion audit log can be stored in a relational database, graph database, vector database, or file system. The system can set different alignment thresholds, conflict thresholds, scoring thresholds, candidate conversion rule thresholds, and deweighting retention thresholds according to specific tasks. Each threshold and weight coefficient can be preset manually or updated based on historical task verification results.
[0079] Through the above implementation methods, the present invention can transform entities, operations, and rules in various artificial intelligence paradigm models into a unified entity syntax five-field structure, and on this basis, complete entity alignment, rule conflict identification, rule arbitration, compound rule generation, and fusion reasoning output, so that the collaborative reasoning process of different paradigm models in the same target task has good rule consistency, result traceability, and subsequent auditability.
[0080] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A multi-paradigm AI rule fusion method based on entity grammar, characterized in that, Includes the following steps: S1. Obtain at least two preset artificial intelligence paradigm models for the same target task and target task data. The preset artificial intelligence paradigm model is a model that can output rule records or can be converted to generate rule records. The rule records include at least rule antecedents and rule consequents. S2. Extract basic data units, abstract combined objects, operation functions, constraint rules and initial task configurations from each preset artificial intelligence paradigm model, and write the extraction results into the corresponding entity syntax five-field structure; S3. Normalize the entity fields and rule fields in the five-field structure of different entity syntaxes to form a cross-normalization entity index table and a normalization rule set; S4. Based on the normalized rule set, perform rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models to obtain cross-paradigm rule conflict results. S5. Based on the cross-paradigm rule conflict results, the conflict rule records are arbitrated to obtain the retention rules, the downgraded rules, and the candidate rules. S6. In the normalized rule set after arbitration, identify the rule combinations that satisfy the cross-paradigm association conditions, and generate cross-paradigm composite rules based on the rule combinations. S7. Write the retention rule, weight reduction rule and cross-paradigm composite rule into the fusion rule set, and perform reasoning on the target task data based on the fusion rule set, and output the fusion reasoning result and the rule source record, conflict handling record and derivation path record corresponding to the fusion reasoning result.
2. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, The preset artificial intelligence paradigm model includes at least two of the following: knowledge graph reasoning model, neural symbol reasoning model, causal inference model, probabilistic inference model, multimodal feature fusion model, and meta-learning model; Specifically, for knowledge graph reasoning models, entities, relations, triples, graph traversal operations, and ontology reasoning rules are extracted; for neural symbolic reasoning models, neural feature vectors, symbolic propositions, mapping functions between neural features and symbolic propositions, and symbolic reasoning rules are extracted; for causal inference models, causal variables, causal edges, intervention operations, observation operations, and causal constraint rules are extracted; for probabilistic inference models, random variables, conditional probability distributions, sampling operations, probability update operations, and probabilistic reasoning rules are extracted; for multimodal feature fusion models, text features, image features, numerical features, modal alignment functions, and cross-modal consistency rules are extracted; and for meta-learning models, task sets, meta-parameters, task adaptation functions, and cross-task transfer rules are extracted.
3. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, The entity syntax five-field structure includes a basic data unit field, an abstract entity field, an operation function field, a rule field, and an initial configuration field; Specifically, input variables that cannot be further decomposed, feature vectors, atomic propositions, triples, or random variables are written into the basic data unit field; semantic entities, causal chains, compound propositions, probabilistic models, cross-modal semantic objects, or task adaptation objects formed by combining basic data units are written into the abstract entity field; functions used to transform, combine, align, infer, sample, update, or adapt basic data units or abstract entities are written into the operation function field; rule records including rule antecedents, rule consequents, rule sources, rule confidence, and rule application conditions are written into the rule field; and target task data, initial observation data, initial symbolic proposition set, prior distribution, domain knowledge base, or multimodal input data are written into the initial configuration field.
4. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, In step S3, when normalizing the entity fields in the five-field structure of different entity syntaxes, an entity alignment score is calculated based on the entity name, entity type, hierarchical relationship, and entity vector representation. The entity alignment score satisfies the following: in, Indicates the first The entity and the first Entity alignment score between entities Indicates the first The entity and the first Name similarity between entities Indicates the first The entity and the first Type consistency between entities Indicates the first The entity and the first The degree of matching of hierarchical relationships between entities Indicates the first The entity and the first Vector proximity between entities , , , These are the corresponding weight coefficients, and ; When the entity alignment score is greater than the preset alignment threshold, the first... The entity and the first Multiple entities are written to the same cross-paradigm entity index entry.
5. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, In step S4, when performing rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models, the rule conflict strength between two rule records is calculated, and the rule conflict strength satisfies: in, Indicates the first Rule record and number The intensity of rule conflicts between rule records Indicates the first Rule record and number Similarity of rule antecedents between rule records Indicates the first Rule record and number The mutual exclusion of rule consequents between rule records. Indicates the first Rule record and number The degree of overlap in the applicable conditions of the rules between the rule records. , , These are the corresponding weight coefficients, and ; When the rule conflict intensity is greater than the preset conflict threshold, determine the first... Rule record and number These rule records constitute cross-paradigm rule conflicts.
6. The multi-paradigm AI rule fusion method based on entity grammar according to claim 5, characterized in that, The cross-paradigm rule conflict includes at least one of the following: conflict of the same antecedent but different consequent, conflict of the antecedent being contained but the consequent being opposite, conflict of the rule and the consequent being mutually exclusive, conflict of the rule's applicable conditions being overlapping but the reasoning results being inconsistent, and conflict of the causal direction being inconsistent with the statistical correlation direction.
7. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, In step S5, when arbitrating conflict rule records, a rule arbitration score is calculated for each conflict rule record, and the rule arbitration score satisfies the following: in, Indicates the first The rule arbitration score of the conflict rule record. Indicates the first The rule confidence of each conflict rule record. Indicates the first The accuracy of the conflict rules recorded in historical target tasks. Indicates the first Each conflict rule record shows the degree of matching with the current target task scenario. Indicates the first Each conflict rule record corresponds to a rule source weight. Indicates the first The intensity of conflict between a conflict rule record and other conflict rule records. , , , , These are the corresponding weight coefficients; Conflict rule records with the highest arbitration score and whose score difference with other conflicting rule records is greater than the first scoring threshold are identified as retained rules; conflict rule records with arbitration scores lower than retained rules but higher than the candidate threshold are identified as downgraded rules; conflict rule records with a difference between rule arbitration scores less than the second scoring threshold are added to the candidate rule pool; conflict rule records whose processing results cannot be determined by automatic arbitration are generated as pending review records.
8. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, The conflict resolution record includes conflict rule identifier, rule source tag, rule antecedent, rule consequent, rule applicable conditions, rule conflict intensity, rule arbitration score, resolution result, and review status; When the verification accuracy of a candidate rule in the candidate rule pool reaches the threshold for candidate rule conversion in subsequent target task verification, the corresponding candidate rule is written into the fusion rule set; when the weighted rule falls below the weighted retention threshold in a preset number of consecutive target task verifications, the corresponding weighted rule is moved from the fusion rule set into the candidate rule pool.
9. The multi-paradigm AI rule fusion method based on entity grammar according to claim 1, characterized in that, In step S6, generating cross-paradigm composition rules includes: Identify the entity association, causal association, semantic similarity, or task dependency relationship between the rule consequent of the first rule record and the rule antecedent of the second rule record; When the association score corresponding to the entity association, causal association, semantic similarity, or task dependency is greater than the composite rule generation threshold, the first rule record and the second rule record are determined as candidate combined rules. Candidate composite rules are generated based on the candidate combination rules. The composite antecedent of the candidate composite rule includes at least the rule antecedent of the first rule record and the rule antecedent of the second rule record. The composite consequent of the candidate composite rule includes at least the rule consequent of the second rule record. The candidate composite rules are subjected to consistency verification of predecessors and successors, verification of rule applicability conditions, and verification of derivation validity. When the consistency verification of predecessor and successor, the rule applicability verification, and the derivation validity verification all pass, the candidate composite rule is determined as a cross-paradigm composite rule and written into the fusion rule set; A fusion derivation path set is generated based on the fusion rule set. The fusion derivation path set is compared with the single paradigm derivation path set corresponding to each preset artificial intelligence paradigm model to obtain new derivation paths that exist only in the fusion derivation path set. The new derivation path is then associated with the rule record it calls, the rule source tag, the conflict handling record, and the fusion reasoning result.
10. A multi-paradigm AI rule fusion system based on entity grammar, characterized in that, include: The model access module is used to acquire at least two preset artificial intelligence paradigm models for the same target task and target task data. The preset artificial intelligence paradigm models are models that can output rule records or can be converted to generate rule records. The rule records include at least rule antecedents and rule consequents. The component mapping module is used to extract basic data units, abstract combined objects, operation functions, constraint rules and initial task configurations from each preset artificial intelligence paradigm model, and write the extraction results into the corresponding entity syntax five-field structure; The normalization module is used to normalize the entity fields and rule fields in the five-field structure of different entity syntaxes, forming a cross-normalization entity index table and a normalization rule set; The rule conflict handling module is used to perform rule antecedent matching and rule consequent compatibility judgment on rule records from different preset artificial intelligence paradigm models according to the normalized rule set, to obtain cross-paradigm rule conflict results, and to arbitrate the conflicting rule records according to the cross-paradigm rule conflict results to obtain retained rules, reduced weight rules and candidate rules. The composite rule generation module is used to identify rule combinations that satisfy cross-paradigm association conditions in the normalized rule set after arbitration processing, and generate cross-paradigm composite rules based on the rule combinations. The fusion inference output module is used to write the retention rules, weight reduction rules and cross-paradigm composite rules into the fusion rule set, and to perform inference on the target task data based on the fusion rule set, and output the fusion inference result as well as the rule source record, conflict handling record and derivation path record corresponding to the fusion inference result.